Where to go from here

Yokome can be extended and improved in a multitude of ways. Some ideas:

• Extend to more languages

• Improve the language model

• Try out a character-based language model (might work well with Chinese and Japanese, since characters are semantically very rich in those languages)

• Provide personalized example sentences based on a user language proficiency estimation

• Provide images alongside the glosses

• Improve support for sentences spanning multiple HTML elements as well as rotated text

• Performance considerations:

  • Block off stopwords

  • Denormalize the dictionary database

  • Make better use of underlying database optimization techniques (espc. caching)

  • Precompute tokenized sentences / disambiguation results

    • Based on recency (starting from the top of the page)

    • Based on word frequencies in the corpus

    • Based on the estimated proficiency of the learner, expressed as a word-frequency range

    • Based on structural elements (headings, links), text size, color, …

  • Improve mouse pointer localization using a binary search on elements

  • Trade disambiguation accuracy for faster processing: Use windowed inputs to the language model instead of a recurrent neural network

• User interface:

  • Add loading indicators

  • Provide better data on entries

    • All headwords

    • More user-friendly presentation of POS tags

    • Restrictions and notes for glosses

  • Make the Yokome infobox’s style independent from the webpage it is displayed on